Estimating Link Capacity in High Speed Networks?

Ling-Jyh Chen1 , Tony Sun2 , Li Lao2 , Guang Yang2 , M. Y. Sanadidi2 , Mario Gerla2
1
Institute of Information Science, Academia Sinica, Taipei 11529, Taiwan
2
Department of Computer Science, UCLA, Los Angeles, CA 90095, USA

Abstract. Knowledge of bottleneck capacity of an Internet path is critical for ef-

ficient network design, management, and usage. With emerging high speed Inter-
net links, most traditional estimation techniques are limited in providing fast and
accurate capacity estimations. In this paper, we propose a new technique, called
PBProbe, to estimate high speed links. PBProbe is based on CapProbe; however,
instead of solely relying on packet pairs, PBProbe employs a “packet bulk” tech-
nique and adapts the bulk length in order to overcome the well known problem
with packet pair based approaches, namely the lack of accurate timer resolution.
As a result, PBProbe not only preserves the simplicity and speed of CapProbe,
but it also correctly estimates link capacities within a much larger range. Using
analysis, we evaluate PBProbe with various bulk lengths and network configura-
tions. We then perform emulation and Internet experiments to verify the accuracy
and speed of PBProbe on high speed links. The results show that PBProbe is
consistently fast and accurate in the great majority of test cases.

1 Introduction
Estimating the bottleneck capacity of an Internet path is a fundamental research prob-
lem in computer networking; knowledge of such capacity is critical for efficient network
design, management and usage. In the past few years, with the growing popularity of
emerging technologies such as overlay, peer-to-peer (P2P), sensor, grid and mobile net-
works, it is becoming increasingly desirable to have a simple, fast and accurate tool
for capacity estimation and monitoring. To accommodate the diversity in network ar-
rangements, an ideal capacity estimation tool must also be scalable and applicable to a
variety of network configurations.
A number of techniques have been proposed for capacity estimation on generic
Internet paths [1–7]. Among them, CapProbe [5] and Pathrate [3] have been well ac-
cepted as two fast and accurate tools in generic network scenarios. However, CapProbe
is a round-trip estimation scheme that works well only on paths consisting of a sym-
metric bottleneck link. Pathrate, on the other hand, is based on histograms and may
converge slowly when the initial dispersion measurements are not of unimodal. As a re-
sult, CapProbe has difficulty estimating capacities of asymmetric links [8], and Pathrate
performs poorly on wireless links [5].
To address the problems above, specialized capacity estimation tools have been
proposed for specific and emerging network scenarios. For instance, ALBP [9] and
?
This work is co-sponsored by the National Science Council and the National Science Founda-
tion under grant numbers NSC-94-2218-E-001-002 and CNS-0435515.
2

AsymProbe [8] are intended for capacity estimation on asymmetric links, and AdHoc
Probe [10] aims to estimate the end-to-end path capacity in wireless networks. However,
for emerging high speed network links (i.e., gigabit links), recent studies have showed
the standing challenging to estimate high speed link capacity (various system issues)
[11], and a simple, fast and accurate technique is still lacking and desirable.
In this paper, we propose a capacity estimation tool for high speed network links,
called PBProbe. PBProbe is inspired by CapProbe. However, instead of solely relying
on one pair of packets, PBProbe employs the concept of “Packet Bulk” to adapt the
number of probing packets in each sample in accordance to the dispersion measurement.
More specifically, when the bottleneck link capacity is expected to be low, PBProbe uses
one pair of packets as usual (i.e. the bulk length is 1). For paths with high bottleneck
capacities, PBProbe increases the bulk length and sends several packets together, which
enlarges the dispersion between the first and and last packet, to overcome the known
timer resolution problem. As we will discuss in more detail later in the paper, timer
resolution is the main challenge in estimating high capacity link capacities [11, 12].
The rest of the paper is organized as follows. In section 2, we summarize related
work on capacity estimation. In section 3, we present and describe PBProbe. In section
4, we present an analysis of PBProbe and evaluate the speed and accuracy of PBProbe
with Poisson cross traffic. In section 5, we evaluate PBProbe on high speed links in our
emulator testbed as well as on the Internet. Section 6 concludes the paper.

2 Related Work

Previous research on capacity estimation relied either on delay variations among probe
packets as illustrated in Pathchar [4], or on dispersion among probe packets as described
in Nettimer [6] and Pathrate [3]. Pathchar-like tools (such as pchar [2] and clink [1])
have limitations in accuracy and speed as shown in [5] [13]. Moreover, they evaluate
the capacity of a link based on the estimates of previous links along the path, thus
estimation errors accumulate and amplify with each measured link [7].
Dispersion-based techniques suffer from other problems. In particular, Dovrolis’
analysis in [3] showed that the dispersion distribution can be multi-modal due to cross
traffic, and that the strongest mode of such distribution may correspond to either (1)
the capacity of the path, or (2) a “compressed” dispersion, resulting in capacity over-
estimation, or (3) the Average Dispersion Rate (ADR), which is always lower than
capacity. Another dispersion-based tool, SProbe [7], exploits SYN and RST packets of
the TCP protocol to estimate the downstream link capacity, and employs two heuristics
to filter out samples which have experienced cross traffic. However, SProbe does not
work properly when the network is highly utilized [14].
Unlike the above approaches, CapProbe [5] uses both dispersion measurements and
end-to-end delay measurements to filter out the packet pair samples that were distorted
by cross traffic. This method has been shown to be both fast and accurate in a variety
of scenarios. The original implementation of CapProbe uses ICMP packets as probing
packets, and it measures the bottleneck capacity on a round-trip basis. As a result, the
capacity estimate does not reflect the higher capacity link when the path is asymmetric.
Other difficulties are encountered when intermediate nodes employ priority schemes to
 3

delay ICMP packet forwarding (e.g. Solaris operating system limits the rate of ICMP
responses, and it is thus likely to perturb CapProbe measurements) [15].
Recent capacity estimation studies have extended the target network scenarios to
more diverse environments. For instance, Lakshminarayanan et al. have evaluated es-
timation tools of capacity and available bandwidth in the emerging broadband access
networks [16]. Chen et al. extended CapProbe to estimate effective path capacity in ad
hoc wireless networks [10]. In addition, ABLP [9] and AsymProbe [8] have been pro-
posed for capacity estimation on the increasingly popular asymmetric links (e.g. DSL
and satellite links).
Nonetheless, capacity estimation on high speed links remains a challenge. Though
recent studies have verified the accuracy of Pathrate in estimating gigabit links [17],
however, the evaluation was done on an emulator-based testbed, which cannot represent
realistic Internet dynamics. Thus, an experimental evaluation of capacity estimation on
high speed links is still lacking.
In this paper, we propose a novel packet bulk technique for estimating high speed
link capacities, called PBProbe. PBProbe is based on CapProbe, but it probes the bottle-
neck link capacity using UDP packets (instead of the ICMP packets used by CapProbe).
We present the PBProbe algorithm in the next section.

3 Proposed Approach: PBProbe

In this section we introduce PBProbe. Similar to CapProbe, PBProbe estimates the link
capacity by actively sending a number of probes to the network and using the minimum
delay sum filter to identify the “good” sample. However, instead of employing a packet
pair, PBProbe uses packet bulk of length k in each probing and measures the capacity
for each direction separately. Specifically, there are two phases in PBProbe. In the first
phase, PBProbe estimates the capacity of the forward link; whereas in the second phase,
PBProbe estimates the capacity of the backward link. Fig. 1 illustrates the algorithm of
PBProbe.
In the first phase (as shown in Fig. 1-a), host A first sends a START packet to host
B to initiate the estimation process. Once the process is initiated, B sends a Request To
Send (RTS) packet to A every G time units. Upon the receipt of the RTS packet, host
A immediately sends B a packet bulk of length k (note: bulk length = k means that
k + 1 packets are sent back to back). For the i-th probing sample, suppose B sends the
RTS packet at time tsend (i) and receives the j-th packet (in the i-th sample) at time
trcv (i, j). The delay sum (i.e. Si ) and the dispersion (i.e. Di ) of the i-th packet bulk
sample are given by:

Si = (trcv (i, 1) − tsend (i)) + (trcv (i, k + 1) − tsend (i)) (1)

Di = trcv (i, k + 1) − trcv (i, 1) (2)

If none of the k + 1 probing packets experience cross-traffic induced queueing, the

sample will reflect the correct capacity. Thus, the “good” sample (say, the m-th sample)
4

Fig. 1. Illustration of PBProbe (a) Phase I: measuring forward direction link capacity; (b) Phase
II: measuring backward direction link capacity.

is identified by applying the minimum delay sum filter to all probing samples (say, n
samples):

m = arg min Si (3)

i=1...n

Therefore, the capacity estimate is made by using the dispersion of the m-th sample
with the minimum delay sum:
kP
C= (4)
Dm
where P denotes the packet size of each probing packet. Since the packet bulk
samples are delivered only in the forward direction, the estimated capacity corresponds
to the bottleneck in this direction.
Once the first phase ends, B sends an END packet to A, and PBProbe enters the
second phase (as shown in Fig. 1-b). In this phase, A first sends an RTS packet (every G
time units) to B, and B replies a packet bulk of length k right upon the receipt of each
RTS packet. Similar to the first phase, PBProbe measures the delay sum and dispersion
for each sample, and estimates the capacity in the backward direction by using the
minimum delay sum filter.

3.1 The Inter-sample Period: G

PBProbe probes the link capacity by sending a packet bulk every G time units. The
value of G is critical for the convergence time of PBProbe. The larger G is, the slower
PBProbe estimation becomes. However, G can not be too small either. If it is too small,
PBProbe is more likely to create congestion in the networks and thus requires longer
time to converge. Therefore, for PBProbe, we set G to be:
2Dm0
G= (5)
U
 5

Fig. 2. The algorithm for determining the appropriate bulk length, k, in PBProbe.

where Dm0 is the dispersion of the good sample (i.e., Dm0 = kP C ), which has the
minimum delay sum among all probing samples seen so far, and U is the maximum
network utilization allowed for PBProbe estimation. We provide a short proof below
showing that if G = 2DUm0 , the network utilization constraint, U , can be guaranteed.

Proof. Since k ≥ 1, we know that

2Dm0 Dm0 Dm0 kP P (k + 1)P

G= ≥ + = + = (6)
U U kU CU CU CU

(k + 1)P
⇒ ≤ CU (7)
G
Let R denote the data rate of the introduced packet bulk probes , i.e., R =
(k+1)P
G ; therefore we can conclude R ≤ CU , i.e., the probing data rate is
never larger than the load constraint, U .

3.2 The Packet Bulk Length: k

The major difference between CapProbe and PBProbe is that PBProbe sends packet
bulks. The purpose of using packet bulks is to overcome the limited system timer reso-
lution, as well as to avoid the additional latency caused by segmentation and reassembly
when the packet size used is larger than the MTU. The algorithm in Fig. 2 is employed
by PBProbe to automatically determine the proper bulk length, k, for capacity estima-
tion.
In the beginning, k is initialized to 1, i.e., PBProbe behaves like CapProbe, us-
ing packet-pair to probe the link capacity. However, whenever the measured disper-
sion is smaller than a certain threshold, say Dthresh , this algorithm will increase the
bulk length (k) by ten-fold and restart the estimation process. Clearly, the decision of
Dthresh value depends on the system timer resolution. In this work, we set Dthresh =
1ms for all the experiments
6

3.3 The Number of Samples: n

CapProbe employs a sophisticated convergence test to determine whether a good sam-
ple has been obtained. To simplify the implementation, PBProbe simply estimates link
capacities using a fixed number of n samples. Obviously, the larger n is, the more ac-
curately PBProbe estimates. The required time for one PBProbe capacity estimate is
linearly proportional to the value of n. More specifically, the larger n is, the longer time
PBProbe requires. Based on the experimental results reported in the previous CapProbe
studies [18] [5], we decide to set n = 200 throughout this paper.

4 Analysis
In this section, we present a queueing model that predicts the probability of obtaining
a “good” sample for a single link with Poisson distributed cross traffic. For simplicity,
we assume the probing samples of PBProbe arrive according to a Poisson process so
that they take so to speak “a random look” at the link. We also assume that the probing
samples do not constitute a significant load on the network since they are sent infre-
quently. Finally, we assume the buffers are large enough so that probing packets will
not be dropped due to buffer overflow. The analytical model is described next, followed
by the results.
Suppose the arrival and service rate of Poisson cross traffic are λ and µ, the service
time of one single probing packet is τ , and the bottleneck link utilization is ρ. The
probability of the first probing packet arriving to an empty system, i.e. p, is given by:
λ
p=1− =1−ρ (8)
µ
Since there is no queueing delay experienced by any probing packets of a “good”
sample, the probability of no queueing delay for the remaining k probing packets (i.e.
no cross traffic packets arrive in the kτ period) is e−λ(kτ ) . Therefore, the probability of
obtaining a “good” sample, i.e. p0 , is given by:

p0 = peλ(−kτ ) = (1 − ρ) e−kλτ (9)

The expected number of samples, N , for obtaining a good sample is then derived
as:
∞
X 1 ekλτ
N= np0 (1 − p0 )n−1 = = (10)
n=1
p0 1−ρ
Suppose the size of the probing packets and the cross traffic packets are equal, then
τ = µ1 = λρ . Therefore, N could be rewritten as:

ekρ
N= (11)
1−ρ
The relationship between the expected number of required samples for one good
sample (N ) and link utilization (ρ) with different packet bulk length (k) is shown in
Fig. 3.
 7

Fig. 3. The expected number of samples (N ) with different link utilization (ρ) and packet bulk
length (k) under Poisson cross traffic.

From the results illustrated in Fig. 3, when k = 1 (i.e. packet-pair based CapProbe),
N is around 25 when the utilization (ρ) is as high as 0.9. However, as k increases, N
increases exponentially. For instance, when ρ = 0.33 , N is around 30 when k = 10,
but becomes around 5,000,000 when k = 50. It turns out that the estimation speed of
PBProbe (i.e. the required number of samples for obtaining a good sample) is highly
related to the employed packet bulk length. Though employing a large packet bulk can
improve the accuracy of dispersion measurements, such large bulk will need a much
larger number of tries in order to lead to a good sample and therefore will slow down
the estimation considerably.

5 Evaluation

5.1 Emulation Experiments

The emulator-based experiments were run on our laboratory testbed. In the scenario,
three testing machines are connected serially with 1 Gbps links. The NISTNet emulator
[20] is installed on the middle machine and can configure the middle machine to create
bottleneck link on the gigabit path. The purpose of this set of experiments is to verify
the needs of bulk length adaption for estimating high speed links; therefore, we used
two fixed bulk lengths (i.e., k = 10 and 100) and did not employ cross traffic in the
emulation experiments. We varied the bottleneck link capacity, and conducted 30 runs
of the experiments for each capacity setting. The results of k = 10 and 100 are shown
in Fig. 4.
Since no cross traffic was present in these experiments, packets within a packet
bulk are expected to traverse the path back-to-back without being disturbed. Therefore,
ideally, the measured dispersion of each packet bulk should represent the undistorted
dispersion corresponding to bottleneck link capacity. Moreover, based on these perfect
3
The utilization of Abilene backbone network, which is the gigabit backbone of Internet2, is
hardly over 30% [19].
8

(a) (b)

Fig. 4. PBProbe estimation results using NISTNet emulator: (a) normalized capacity estimates
and (b) coefficient of variation of capacity estimates with various bottleneck capacity settings.

Table 1. Required system timer resolution for accurate PBProbe estimation (assuming probing
packet size is 1500 bytes).

Bottleneck Link Capacity

k 1Gbps 100Mbps 10Mbps
1 0.012ms 0.12ms 1.2ms
10 0.12ms 1.2ms 12ms
100 1.2ms 12ms 120ms

dispersions, the capacity estimate should be accurate and consistent for every probing
sample.
However, since PBProbe requires accurate dispersion measurements, the capacity
estimates tend to become inaccurate when timer resolution is inadequate. Table 1 shows
the required timer resolution of PBProbe on links with different capacity and for differ-
ent bulk lengths. Obviously, when the bottleneck link capacity is high or the bulk length
is small, a high timer resolution is required. For instance, with link capacity = 100 Mbps
and packet pairs (i.e., k = 1), only a powerful processor can satisfy the required resolu-
tion of 0.12 ms. As the capacity increases to 1 Gbps, the required 0.012 ms resolution
becomes very difficult to achieve, and the measurement accuracy will degrade. Indeed,
this trend was observed in Fig. 4-a and 4-b. In these two figures, for small k, the capac-
ity estimates become increasingly inaccurate and inconsistent when the required timer
resolution became greater (or equivalently, when the bottleneck capacity was enlarged).
However, the results also reveal that, when measuring high capacity links, a large k
can successfully alleviate the inaccuracy and inconsistency of capacity estimates caused
by poor timer resolution. For instance, when the bottleneck link capacity is 600 Mbps,
PBProbe with bulk length k = 100 can accurately estimate the capacity (as illustrated
by very small coefficient of variation on capacity estimates). In contrast, when k =
10, PBProbe can only measure around 60% of the link capacity, and the coefficient of
variation is much larger. Based on the experimental results as well as the analysis shown
 9

Fig. 5. Topology and path properties (bottleneck capacity and round trip delay) of selected gigabit
Internet paths.

in Table 1, we conjecture that our testbed hosts can only provide timer resolution of
approximately 1 ms. Therefore, we decided to fix Tthresh = 1ms in the k adaptation
algorithm of PBProbe for all the following experiments.

5.2 Internet Experiments

Here, we conducted Internet experiments to evaluate PBProbe in realistic environments.

Five Internet hosts (CalTech: California Institute of Technology; GaTech: Georgia In-
stitute of Technology; NTNU: National Taiwan Normal University; PSC: Pittsburgh
Supercomputing Center; UCLA: University of California at Los Angeles) and five In-
ternet gigabit paths (within California: UCLA and CalTech); across country: UCLA -
PSC, PSC - GaTech, GaTech - UCLA; and international: NTNU - UCLA) were se-
lected for the experiments. The topology and path properties of the selected paths are
illustrated in Fig. 5.
Fig. 6 and 7 illustrate the experiment results (i.e. normalized mean and coefficient
of variation of capacity estimates in 20 runs). It is clear that, in Fig. 6, the normalized
capacity estimates are mostly within 90% accuracy range, except three outgoing links
from UCLA to CalTech, GaTech, and PSC. This is due to the fact that the outgoing
link of UCLA backbone has around 30% utilization, which is much higher than the
utilization of the incoming link (around 15%) [21], and, in this case, PBProbe requires
a large number of samples in order to correctly estimate the capacity. Since we set
n = 200 in all experiments, PBProbe only estimated around 80% of the link capacity.
In addition, Fig. 7 also shows that the coefficient of variation of PBProbe estimates
are below 0.15 on all tested links, i.e. the capacity estimates (20 runs) of each high-
speed link are very stable. Comparing with the results shown in Fig. 4, it turns out that
PBProbe was able to adapt its bulk length to the most appropriate value (i.e. k = 100)
so that it could consistently and accurately estimate the capacities of high speed links.
10

Fig. 6. PBProbe experiment results (mean of 20 runs) on high speed links.

Fig. 7. PBProbe experiment results (coefficient of variation of 20 runs) on high speed links.

Table 2. Comparison of PBProbe and Pathrate Table 3. Comparison of PBProbe and Pathrate
on Internet links. (Capacity: Mbps; Time: sec) overhead on Internet gigabit links.

PBProbe Pathrate GaTech → UCLA UCLA → GaTech

Capacity Time Capacity Time PBProbe Pathrate PBProbe Pathrate
CalTech → UCLA 919.6 14 933.2 17 spent time 14 sec 18 sec 14 sec 1223 sec
UCLA → CalTech 839.4 14 945.3 1146 total packets 20,213 2,414 20,213 27,630
GaTech → UCLA 928.1 14 932.9 18 total bytes 30,319,500 3,543,752 30,319,500 39,707,740
UCLA → GaTech 840.3 14 968.1 1223 BW consumption 2.166Mbps 1.575Mbps 2.166Mbps 0.260Mbps

PSC → GaTech 959.9 13 995.4 1122

GaTech → PSC 905.9 13 947.0 17
PSC → UCLA 889.6 14 935.6 20
UCLA → PSC 845.2 15 905.6 20
NTNU → UCLA 580.6 20 575.6 1641
UCLA → NTNU 588.4 21 573.4 1641

5.3 Comparisons of PBProbe and Pathrate

So far, we have evaluated PBProbe in a variety of network scenarios. In this subsection,

we compare the performance of PBProbe and Pathrate in terms of accuracy, speed, and
bandwidth consumption (i.e. overhead). The experiments are performed on the same set
of high speed Internet links as illustrated in Fig. 5, and the results of capacity estimates
and required time are shown in Table 2.
In Table 2, PBProbe measures at least around 85% bottleneck capacities of all tested
links; whereas Pathrate is very accurate and measures at least 90% bottleneck capacities
for all links. However, for some links with long delay and/or high utilization, Pathrate
required more than 1000 seconds to estimate the capacity. This is due to the fact that,
if the distribution of measured dispersion is not unimodal after the first phase, Pathrate
will start the second phase to probe the network using different packet train lengths and
 11

packet sizes. Once Pathrate enters the second phase, it takes long time to determine the
correct link capacity. As a result, Pathrate converges fast if the dispersion distribution
is unimodal in the first phase, but it becomes much slower than PBProbe otherwise.
We also compared the packet overhead caused by PBProbe and Pathrate on high
speed Internet links. Table 3 shows the comparison on one of the high speed links, the
UCLA - GaTech link. From the experiment results, PBProbe is more expensive than
Pathrate, if Pathrate only uses one phase to estimate link capacity on high speed links.
In case when Pathrate is required to enter the second phase, PBProbe and Pathrate
produce a comparable amount of packet overhead. However, since Pathrate is much
slower after entering the second phase, the bandwidth consumption (i.e. bits per second)
is much smaller than PBProbe. It is worth pointing out that, even though the bandwidth
consumption of PBProbe (approximately 2 Mbps) seems to be relatively high, it is
realistically only 0.2% of the bottleneck capacity (i.e., 1 Gbps); thus, it is not intrusive
to other traffic flows in the network.
The experiment results suggest that there are trade-offs between PBProbe and Pathrate
for high-speed path capacity estimation. On the one hand, PBProbe yields very good
estimation results rapidly (e.g., less than 20 seconds in most cases). If given more time,
it will progressively improve the estimates, since better samples can be obtained. On the
other hand, Pathrate tends to produce accurate results, but the required time may vary
from approximately 20 seconds to 20 minutes. Therefore, Pathrate may not be ideal in
scenarios when an estimation of the bottleneck capacity needs to be obtained within a
very short time.
It should also be noted that the packet overhead of PBProbe is proportional to the
employed bulk length k. While measuring a high speed link, PBProbe increases its
bulk length and in turn increases the packet overhead, in order to overcome the limited
support of system timer resolution. Nonetheless, thank to the employed U parameter,
the bandwidth consumption of PBProbe is restricted by the utilization upper bound.
Hence, PBProbe can carefully control the trade-off between the bandwidth consumption
and the required time in order to satisfy the requirement of different applications.

6 Conclusions

In this paper, we studied a classic problem of link capacity estimation, and we proposed
a technique, called PBProbe, to estimate bottleneck capacity for emerging high speed
links. PBProbe is based on the CapProbe algorithm, but it uses “packet bulk” to adapt
the number of packets in each probing according to different network characteristics. As
a result, it preserves the simplicity, speed, and accuracy of CapProbe, as well as over-
coming the poor system timer resolution problem on high speed links. Using analysis,
emulation, and Internet experiments, we evaluated the accuracy, speed, and overhead
of PBProbe on various network configurations. The results show that PBProbe can cor-
rectly and rapidly estimate bottleneck capacity in almost all test cases. Comparing to
other capacity estimation techniques, PBProbe is ideal in real deployments that requires
online and timely capacity estimation. This capacity estimation technique can further
provide assistance to typical applications such as peer-to-peer streaming and file shar-
12

ing, overlay network structuring, pricing and QoS enhancements, as well as network
monitoring.

7 Acknowledgments
We are grateful to the following people for their help in carrying out PBProbe mea-
surements: Sanjay Hegde (CalTech), Che-Chih Liu (NTNU), Cesar A. C. Marcondes
(UCLA), and Anders Persson (UCLA).

References
1. “Clink: a tool for estimating internet link characteristics,”
http://allendowney.com/research/clink/.
2. “pchar: A tool for measuring internet path characteristics,”
http://www.kitchenlab.org/www/bmah/Software/pchar/.
3. C. Dovrolis, P. Ramanathan, and D. Moore, “What do packet dispersion techniques mea-
sure?” in IEEE Infocom, 2001.
4. V. Jacobson, “Pathchar: A tool to infer characteristics of internet paths,”
ftp://ftp.ee.lbl.gov/pathchar/. [Online]. Available: ftp://ftp.ee.lbl.gov/pathchar/
5. R. Kapoor, L.-J. Chen, L. Lao, M. Gerla, and M. Y. Sanadidi, “Capprobe: A simple and
accurate capacity estimation technique,” in ACM SIGCOMM, 2004.
6. K. Lai and M. Baker, “Measuring bandwidth,” in IEEE Infocom, 1999, pp. 235–245.
7. S. Saroiu, P. K. Gummadi, and S. D. Gribble, “Sprobe: A fast technique for measuring bot-
tleneck bandwidth in uncooperative environments,” in IEEE Infocom, 2002.
8. L.-J. Chen, T. Sun, G. Yang, M. Y. Sanadidi, and M. Gerla, “End-to-end asymmetric link
capacity estimation,” in IFIP Networking, 2005.
9. Y. Lin, H. Wu, S. Cheng, W. Wang, and C. Wang, “Measuring asymmetric link bandwidths
in internet using a multi-packet delay model,” in IEEE ICC, 2003.
10. L.-J. Chen, T. Sun, G. Yang, M. Y. Sanadidi, and M. Gerla, “Adhoc probe: Path capacity
probing in ad hoc networks,” in WICON, 2005.
11. G. Jin and B. Tierney, “System capability effect on algorithms for network bandwidth mea-
surement,” in ACM IMC, 2003.
12. R. Kapoor, L.-J. Chen, M. Y. Sanadidi, and M. Gerla, “Accuracy of link capacity estimates
using passive and active approaches with capprobe,” in IEEE ISCC, 2004.
13. K. Lai and M. Baker, “Measuring link bandwidths using a deterministic model of packet
delay,” in ACM SIGCOMM, 2000.
14. S.-J. Lee, P. Sharma, S. Banerjee, S. Basu, and R. Fonseca, “Measuring bandwidth between
planetlab nodes,” in PAM, 2005.
15. S. Savage, “Sting: a tcp-based network measurement tool,” in USENIX Symposium on Inter-
net Technologies and Systems, 1999.
16. K. Lakshminarayanan, V. N. Padmanabhan, and J. Padhye, “Bandwidth estimation in broad-
band access networks,” in IMC, 2004.
17. R. Prasad, M. Jain, and C. Dovrolis, “Evaluating pathrate and pathload with realistic cross-
traffic,” http://www.cc.gatech.edu/ jain/pub/talk/best03 talk.ppt, 2003 Bandwidth Estimation
Workshop. [Online]. Available: http://www.cc.gatech.edu/ jain/pub/talk/best03 talk.ppt
18. L.-J. Chen, T. Sun, D. Xu, M. Y. Sanadidi, and M. Gerla, “Access link capacity monitoring
with tfrc probe,” in E2EMON, 2004.
19. “Abilene network traffic,” http://loadrunner.uits.iu.edu/weathermaps/abilene/.
20. “Nistnet: network emulation package,” http://www.antd.nist.gov/itg/nistnet/.
21. “Cenic network statistics,” http://cricket.cenic.org/grapher.cgi.